Lenny's Podcast · Published 2026-09-28

Roles aren't converging—they're expanding | Tamar Yehoshua (Atlassian CPO)

privateReview: unreviewedConfidence: high
ai buildersproduct managementrole expansionTeamwork Grapheval loopsai-builderseval-harnesscontext-graphMCProle-expansionagentic-workflowsvelocity-metricsvibecoding
Open on YouTube ↗

Summary

Overview

  • Speaker: Tamar Yehoshua
  • Channel: Lenny's Podcast
  • Main topic: How AI tools are expanding the role of Product Managers (PMs) rather than eliminating it, with practical examples from Atlassian.
  • Purpose: To explain how PM roles are evolving in the era of AI builders, share concrete examples of Atlassian PMs leveraging AI tools, and provide a framework for measuring AI fluency and team acceleration. Tamar Yehoshua, Chief Product and AI Officer at Atlassian, shares insights from the Lenny & Friends Summit on September 10, 2026. Addressing fears that AI would eliminate PMs, engineers, or designers, she argues that roles aren't converging or disappearing; instead, they are expanding. PMs now have a front-row seat to AI acceleration, doing more work across prototyping, writing evals, going deeper on customer feedback, codifying best practices, and consuming information at velocity using Atlassian's Teamwork Graph and Rovo.

Topic Map

The AI Builder Era and Role Expansion

  • Explanation: Discusses the shift from fear of job loss to the reality of role overlap and expansion, where founders and PMs wear multiple hats with AI tools.
  • Key claims:
    • Instead of roles going away, jobs are overlapping in the AI builder era.
    • Founders have always been builders wearing many hats; now PMs and teams can do the same.
    • The job of a PM remains fundamentally the same—finding product-market fit, building products people love, and building viable businesses—while HOW it is done has changed.
  • Examples:
    • Startups hiring builders rather than siloed PMs or designers.
    • Atlassian PMs checking in code, writing evals, and fixing design bugs.
  • Terminology:
    • AI builder
    • role expansion
    • product-market fit
  • Why it matters: Understanding role expansion prevents unnecessary fear and empowers PMs to leverage AI tools to accelerate delivery and customer value.

Three Concrete Atlassian Examples

  • Explanation: Detailed case studies from Atlassian showing how PMs expanded their roles using AI tools across Confluence, RovoClaw, and Jira.
  • Key claims:
    • Confluence team: A PM with zero prior coding experience checked in 26 PRs in a single month by partnering with engineers on harnesses, resulting in 2x throughput in evals.
    • RovoClaw team: PM and designer vibecoded an alpha from scratch, proving the value before bringing in engineers, then shifting from rowing to steering to unblock the team.
    • Jira team: PMs and engineers built AI features in Jira to reinvent the SDLC, resulting in 3x throughput and 22 user-facing features shipped in 10 weeks.
  • Examples:
    • Remix with Rovo and Confluence slides launching in 6 weeks instead of 6 months.
    • Using Figma MCP and coding agents to fix 14 design bugs in an hour.
  • Terminology:
    • vibecoding
    • eval loops
    • Figma MCP
    • coding agent
  • Why it matters: Demonstrates practical, real-world applications of AI tools in enterprise software development and product management.

The AI Fluency Index Framework

  • Explanation: Atlassian's internal framework to train and assess PMs on AI capabilities to help them become AI-native.
  • Key claims:
    • The AI Fluency Index assesses 6 capabilities from 1 (curious) to 5 (pioneering).
    • Capabilities include tool usage, writing evals, automating data insights, prototyping, and technical literacy.
    • AI Builders Weeks (held once a quarter) train PMs and designers with external speakers, internal leaders, and hands-on projects.
  • Examples:
    • Claire speaking at Atlassian's Builders Week.
    • Over 1,000 participants building 120 new workflows during Builders Week.
  • Terminology:
    • AI Fluency Index
    • AI Builders Week
    • technical literacy
    • curious vs pioneering
  • Why it matters: Provides organizations with a structured way to upskill PMs and measure their progress toward AI fluency.

Key Points

Roles are expanding, not converging

  • Explanation: While people initially worried AI would eliminate PMs or engineers, everyone can now do much more across domains.
  • Evidence: PMs checking in code, writing evals, fixing design bugs, and prototyping.
  • Practical implication: PMs must embrace adjacent skills like prototyping and eval writing to remain competitive.

From rowing to steering

  • Explanation: PMs initially roll up their sleeves to build and code (rowing), then transition to steering (unblocking, prioritizing, setting direction) as teams scale.
  • Evidence: RovoClaw project where PM started coding the alpha and later shifted to unblocking the team.
  • Practical implication: Knowing when to shift from hands-on building to strategic steering optimizes team velocity.

Context is everything

  • Explanation: AI tools require enterprise context and teamwork graphs to be effective and safe.
  • Evidence: Atlassian's Teamwork Graph provides full context across Jira, Confluence, and Slack.
  • Practical implication: Organizations must integrate tools and data systems so AI agents have proper context.

Frameworks, Models & Processes

AI Fluency Index

  • How it works: Measures PM capabilities across 6 core areas on a 1-5 scale (1=curious, 5=pioneering).
  • Components:
    • Tool usage
    • Writing evals
    • Automating data insights
    • Prototyping
    • Technical literacy
    • Outcome focus
  • When to use: To assess and guide PM skill development toward becoming AI-native.

Examples & Case Studies

A PM with no coding experience checked in 26 PRs in a month on the Confluence team.

  • Illustrates: PM role expansion into code check-ins using engineering partnerships and AI tools.
  • Lesson: PMs can directly contribute to code and accelerate delivery when paired with proper harnesses.

RovoClaw project started with a PM and designer vibecoding an alpha from scratch before adding engineers.

  • Illustrates: 0-to-1 prototyping and transitioning from rowing to steering.
  • Lesson: Prototyping early with AI helps validate ideas before scaling engineering resources.

Jira team used Jira Agent in Slack to triage 900 pieces of customer feedback with 90% accuracy.

  • Illustrates: Customer insights triage and automated feedback loops.
  • Lesson: AI agents can drastically reduce manual research and routing time for engineering teams.

Actionable Takeaways

  • Immediate:
    • Stop believing everything you read on X/Twitter about job replacement.
    • Focus on customer needs and how AI tools can get them value faster.
    • Build prototypes and write evals to accelerate your product workflow.
  • Strategic:
    • Embrace role expansion and learn technical skills as a PM.
    • Implement AI Builders Weeks to continuously upskill teams.
    • Leverage enterprise context graphs (like Atlassian's Teamwork Graph) for AI tools.
  • Questions to investigate:
    • How are you measuring developer and PM velocity and outcomes in the AI era?
    • What is the right balance of rowing vs. steering for PMs on your team?
    • How can your organization establish an AI fluency index tailored to your products?

Claims Worth Verifying

  • Confluence team shipped Remix and Confluence slides in 6 to 8 weeks instead of 6 months using AI and tighter eval loops. (Case study metric)
  • Jira team achieved 3X throughput and shipped 22 user-facing features in 10 weeks. (Productivity claim)
  • Jira Service Management agent triaged 900 pieces of customer feedback with 90% accuracy. (AI accuracy metric)

Notable Quotes

"Roles aren't converging, they're expanding." "Everything that you read on X, don't believe." "How do you stay on top of it all? And the answer is you don't."

Compressed Summary

  • AI tools expand PM roles rather than replacing them.
  • Atlassian case studies show 2x to 3x throughput improvements across Confluence, RovoClaw, and Jira.
  • The AI Fluency Index helps train and assess PM skills across 6 capabilities.
  • PMs shift from rowing (coding, prototyping) to steering (prioritizing, unblocking) as projects mature.
  • Keywords: ai builders, product management, role expansion, teamwork graph, eval loops
  • Core insight: AI tools do not eliminate PM roles; they expand them, allowing PMs to prototype, write evals, and accelerate delivery by shifting from rowing to steering.

Core insights

7
Mechanismhigh noveltymoderate evidence

A product manager with no prior coding experience checked in 26 PRs in a single month by pairing with engineers on purpose-built 'harnesses' — implying the harness/scaffolding, not the person's coding ability, is the enabling abstraction that lets non-engineers safely produce production code.

Why it matters

Reframes the bottleneck for expanding contributor pools from individual skill to harness design: if you invest in scaffolding, guardrails, and eval loops, non-engineer domain experts can contribute code directly. This argues for treating harnesses as first-class engineering deliverables rather than throwaway glue.

Generalization

Wherever a task is gated by scarce specialist skill, a well-designed harness can convert domain knowledge into throughput, shifting the design problem to the scaffolding rather than the training of the human.

A PM with zero prior coding experience checked in 26 PRs in a single month by partnering with engineers on harnesses, resulting in 2x throughput in evals.
Open source video
Practicehigh noveltymoderate evidence

Vibecoded prototypes are being used as a pre-engineering validation step: a PM and designer built an alpha from scratch with AI, proved the value, and only then brought in engineers — then the PM shifted from 'rowing' (building) to 'steering' (unblocking, prioritizing, direction-setting).

Why it matters

Suggests a new team formation pattern where prototype validity is established before engineering headcount is committed, and where an individual's mode of contribution is expected to change over the project lifecycle rather than stay fixed. Architecturally this implies prototypes are explicitly disposable and the handoff point must be designed.

Generalization

AI-assisted prototyping can serve as a cheap gate before expensive specialist allocation, but introduces an explicit handoff boundary where the artifact is either promoted or discarded.

RovoClaw team: PM and designer vibecoded an alpha from scratch, proving the value before bringing in engineers, then shifting from rowing to steering to unblock the team.
Open source video
Architecturemedium noveltyweak evidence

Agent effectiveness is framed as dependent on an enterprise context graph spanning tools (Jira, Confluence, Slack) — Atlassian's Teamwork Graph — rather than on model capability alone. Context is the substrate, not a feature.

Why it matters

Places the context/integration layer as the central architectural asset for agentic systems in enterprises. If correct, competitive differentiation lives in cross-tool context assembly and permissioning, not in the agent loop itself.

Generalization

For agents operating in real organizations, the cross-system context graph (identity, permissions, history across tools) is likely a more durable moat than prompt or model choice.

Atlassian's Teamwork Graph provides full context across Jira, Confluence, and Slack.
Open source video
Organizations must integrate tools and data systems so AI agents have proper context.
Open source video
Mechanismmedium noveltymoderate evidence

Design bugs were fixed in an hour by routing Figma through an MCP server into coding agents — a concrete instance of MCP acting as the seam between a domain tool of record and an agent runtime.

Why it matters

Shows a repeatable integration pattern: expose a specialist tool via MCP (or equivalent) and let coding agents operate across tool boundaries. This is a concrete, testable integration architecture rather than a slogan about 'connecting tools.'

Generalization

Standardized tool-interface protocols (MCP-style) turn siloed tools into agent-addressable capabilities, collapsing cross-tool workflows into single agent runs.

Using Figma MCP and coding agents to fix 14 design bugs in an hour.
Open source video
Practicemedium noveltymoderate evidence

The AI Fluency Index scores practitioners on six capabilities (tool usage, writing evals, automating data insights, prototyping, technical literacy, outcome focus) on a 1–5 scale from 'curious' to 'pioneering', paired with quarterly AI Builders Weeks (1,000+ participants, 120 new workflows).

Why it matters

Provides a concrete, replicable organizational mechanism for measuring and driving agent-era skill — a rubric plus a recurring hands-on lab — which is a measurable intervention rather than a training mandate. The inclusion of eval-writing and outcome focus as scored competencies is notable.

Generalization

Rubric-based fluency measurement combined with periodic hands-on build events is a transferable pattern for building capability in any rapidly evolving tool domain.

The AI Fluency Index assesses 6 capabilities from 1 (curious) to 5 (pioneering).
Open source video
Over 1,000 participants building 120 new workflows during Builders Week.
Open source video
Empirical Resultmedium noveltymoderate evidence

An agent deployed inside an existing communication surface (Jira Agent in Slack) triaged 900 pieces of customer feedback at a claimed 90% accuracy — embedding the agent where the team already works rather than in a new UI.

Why it matters

Suggests that distribution/adoption for triage-style agents comes from living inside existing surfaces, and that accuracy is the operative reliability metric to report and monitor for feedback-routing agents.

Generalization

For workflow agents, placing the agent in the surface where the work already happens may reduce adoption friction more than building a dedicated interface; accuracy should be tracked as the primary quality signal.

Jira team used Jira Agent in Slack to triage 900 pieces of customer feedback with 90% accuracy.
Open source video
Mental Modelmedium noveltyweak evidence

The framing that roles are expanding rather than converging — the PM job's WHAT (product-market fit, viable businesses) is unchanged while the HOW changes — is used to justify cross-boundary contribution (PMs checking in code, writing evals, fixing design bugs).

Why it matters

Contradicts the common displacement narrative and predicts role overlap rather than role collapse, which affects how teams are structured and what tooling (harnesses, evals, context graphs) is worth investing in.

Generalization

As tool capability rises, expect role boundaries to blur toward overlap rather than specialization, with the underlying objective of a function staying constant while its execution surface widens.

Instead of roles going away, jobs are overlapping in the AI builder era.
Open source video
The job of a PM remains fundamentally the same—finding product-market fit, building products people love, and building viable businesses—while HOW it is done has changed.
Open source video

Deep dives

4

Harness design for non-engineer code contribution

Research question

What minimal harness (scaffolding, eval loop, review gate) enables a domain expert to contribute production code, and how does its design affect PR acceptance and defect rate?

Why

If harnesses, not individual skill, gate contribution, then treating them as first-class engineering artifacts could widen contributor pools and change how teams hire and train.

A PM with zero prior coding experience checked in 26 PRs in a single month by partnering with engineers on harnesses, resulting in 2x throughput in evals.
Open source video
Source video

Cross-tool context graph as agent substrate

Research question

To what extent does agent effectiveness depend on a cross-tool context graph versus model capability, and can such graphs be reproduced outside a single vendor's ecosystem with proper permissions?

Why

If context assembly is the durable moat, architecture effort should shift to integration, permissions, and context freshness rather than model selection.

Atlassian's Teamwork Graph provides full context across Jira, Confluence, and Slack.
Open source video
Organizations must integrate tools and data systems so AI agents have proper context.
Open source video
Source video

Rowing-to-steering transition criteria

Research question

What measurable signals (team size, handoff latency, review load) indicate the optimal point for a builder to shift from rowing (building) to steering (unblocking)?

Why

Mistiming the shift risks bottlenecking the team or losing hands-on credibility; a heuristic would make the pattern transferable.

RovoClaw team: PM and designer vibecoded an alpha from scratch, proving the value before bringing in engineers, then shifting from rowing to steering to unblock the team.
Open source video
Source video

MCP as integration seam for agent workflows

Research question

Does routing specialist tools through MCP-style interfaces measurably reduce cross-tool workflow time and error rate compared to manual handoffs or custom integrations?

Why

If standardized tool interfaces collapse cross-tool workflows into single agent runs, that changes how tools are designed and integrated.

Using Figma MCP and coding agents to fix 14 design bugs in an hour.
Open source video
Source video

Article ideas

4

Harnesses Are the New Hiring Bar: Why Non-Engineers Can Ship Code

The bottleneck for expanding code contribution is not individual coding skill but the quality of harnesses, guardrails, and eval loops; teams should invest in these as core engineering artifacts.

Angle

Argue against training non-engineers to code; show how harness design converts domain expertise into throughput.

Source video

Your Agent's Moat Is the Context Graph, Not the Model

In enterprise settings, agent effectiveness is determined by cross-tool context assembly and permissioning, not by model choice; competitive differentiation shifts to integration layers.

Angle

Contrast model-centric AI strategies with context-centric ones; use Teamwork Graph as evidence.

Source video

Ship Agents Where Work Already Happens

Embedding agents in existing communication surfaces (e.g., Slack) reduces adoption friction more than building a dedicated UI, but forces accuracy to become the headline reliability metric.

Angle

Use Jira Agent in Slack triage case to argue for surface-first agent distribution.

Source video

The Prototype Is the New Gate Before Engineering Headcount

AI-assisted prototyping by PMs/designers can serve as a cheap validation step before committing engineers, but teams need explicit promotion/discard criteria to avoid debt.

Angle

Analyze the RovoClaw pattern and the rowing-to-steering transition; propose handoff design.

Source video

Project ideas

4

Domain-Expert Harness

beyond-evals

A domain expert (non-engineer) using a purpose-built harness (scaffolding + eval loop + review gate) will achieve >80% PR acceptance rate and <5% defect rate on well-scoped tasks.

Proof of concept

Build a minimal harness for a specific domain (e.g., data pipeline config) and have a non-engineer produce 10 PRs over 2 weeks.

Measurement

PR acceptance rate, defect rate, time to merge vs baseline engineer.

Source video

Permission-Aware Context Graph

gatehouse

An MCP-style context aggregation layer across 3+ heterogeneous tools improves agent task accuracy by >30% relative to single-tool context, while enforcing consistent permissions.

Proof of concept

Prototype a context layer that indexes and retrieves from Jira/Confluence/Slack-like tools, with permission-aware retrieval.

Measurement

Agent accuracy on a set of tasks, permission violation rate.

Source video

Rowing-to-Steering Signal Detector

movement-lab

Project signals (team size, handoff latency, review load) can predict the optimal time for a builder to shift from rowing to steering, and timing the shift based on these signals improves team throughput vs fixed schedule.

Proof of concept

Instrument a small team project, collect signals, correlate with outcomes.

Measurement

Correlation between signal threshold and throughput/quality post-shift.

Source video

Triage Accuracy Replication

beyond-evals

The reported 90% triage accuracy is reproducible with a labelled ground truth and residual errors concentrate in specific categories (e.g., ambiguous feedback).

Proof of concept

Replicate feedback triage with a labelled dataset of 900 items, compare agent to human.

Measurement

Accuracy, error distribution by category.

Source video

Architectural implications

3

Non-engineer contribution is gated by harness design rather than by the contributor's raw skill.

Before

Code contribution requires the contributor to hold the engineering skill; harnesses are incidental glue.

After

Harnesses become the deliberate boundary that lets domain experts contribute code safely.

Consequence

Teams invest in eval loops, guardrails, and scaffolding as core engineering artifacts to widen the contributor pool.

Source video

Agent capability is attributed to a cross-tool context graph (Teamwork Graph) spanning Jira, Confluence, and Slack.

Before

Agent quality is treated as a function of model and prompt.

After

Context assembly across systems of record becomes the central differentiating layer.

Consequence

Architecture effort and competitive moat shift toward integration, permissions, and context graphs.

Source video

Agents are deployed inside existing surfaces (Jira Agent in Slack) rather than in bespoke UIs.

Before

New AI capability ships as a new standalone interface.

After

Capability ships where the workflow already lives.

Consequence

Reduced adoption friction, but the agent inherits the surface's constraints and its accuracy becomes the headline reliability metric.

Source video

Tradeoffs and failure modes

4

Rowing vs steering transition for PMs

Benefit

PMs who build hands-on validate ideas fast and unblock teams by understanding the work deeply.

Cost or risk

No stated mechanism for deciding when to stop rowing and start steering; mistiming risks either bottlenecking the team or losing hands-on credibility.

then shifting from rowing to steering to unblock the team.
Open source video
Source video

Vibecoded alpha before engineers join

Benefit

Proves value cheaply before committing engineering headcount.

Cost or risk

A PM/designer-built alpha may not be a viable base for production; the summary does not state how the prototype is promoted or discarded, leaving a potential debt/rewrite risk.

PM and designer vibecoded an alpha from scratch, proving the value before bringing in engineers
Open source video
Source video

Throughput claims from AI-enabled workflows

Benefit

Large reported gains (2x eval throughput, 3x throughput, 22 features in 10 weeks, 6 weeks instead of 6 months).

Cost or risk

Claims are single-source, self-reported vendor metrics with no stated measurement method or baseline; cannot be treated as established results.

resulting in 3x throughput and 22 user-facing features shipped in 10 weeks
Open source video
Source video

Organizational upskilling via AI Builders Weeks

Benefit

1,000+ participants produced 120 new workflows, giving broad hands-on exposure.

Cost or risk

Quarterly cadence and large participant counts require sustained investment; the summary gives no evidence of durable retention or impact on shipped outcomes.

AI Builders Weeks (held once a quarter) train PMs and designers with external speakers, internal leaders, and hands-on projects.
Open source video
Source video

Open questions

5

What concretely constitutes the 'harness' that let a non-coding PM land 26 PRs, and can it be specified as a reusable pattern?

Why unresolved

The summary names harnesses as the enabling factor but does not describe their structure, review gates, or eval integration.

Research direction

Design and benchmark a minimal harness (scaffolding + eval loop + review gate) that lets a domain expert contribute code, measuring PR acceptance rate and defect rate.

Source video

What is the correct trigger or heuristic for a builder to shift from rowing to steering?

Why unresolved

The summary asserts the transition happened on RovoClaw but gives no criterion for timing it.

Research direction

Instrument projects to identify signals (team size, handoff latency, review load) that correlate with the optimal rowing-to-steering switch point.

Source video

How should PM/developer velocity and outcome be measured in the AI era?

Why unresolved

The summary lists this as an open question posed to the audience rather than answered.

Research direction

Compare throughput metrics (PRs, features shipped) against outcome metrics (customer value, retention) to determine which are robust under AI-assisted workflows.

Source video

Can the cross-tool context graph (Teamwork Graph) be reproduced outside a single vendor's ecosystem, and what permissions/security model does it require?

Why unresolved

The summary states the graph provides context across Jira, Confluence, and Slack but gives no architecture, permissioning, or portability detail.

Research direction

Prototype an MCP-style context aggregation layer across heterogeneous tools and evaluate permission handling and context freshness.

Source video

How is the 90% triage accuracy of the Jira Agent in Slack measured and what failure modes remain in the other 10%?

Why unresolved

The summary reports the accuracy figure without methodology, ground-truth source, or error analysis.

Research direction

Replicate feedback-triage evaluation with labelled ground truth and characterize the residual error distribution by category.

Source video

Key claims

8
factualVerification needed

A PM with no prior coding experience checked in 26 PRs in a single month by partnering with engineers on harnesses.

Evidence

A PM with zero prior coding experience checked in 26 PRs in a single month by partnering with engineers on harnesses

Question

What did the harness consist of, what review gates applied, and were all 26 PRs merged to production?

Source video
causalVerification needed

Harness-mediated PM contribution resulted in 2x throughput in evals on the Confluence team.

Evidence

resulting in 2x throughput in evals

Question

Against what baseline and over what period was the 2x eval throughput measured?

Source video
comparativeVerification needed

The Jira team achieved 3x throughput and shipped 22 user-facing features in 10 weeks by building AI features in Jira to reinvent the SDLC.

Evidence

PMs and engineers built AI features in Jira to reinvent the SDLC, resulting in 3x throughput and 22 user-facing features shipped in 10 weeks

Question

What definition of throughput was used, and what was the pre-intervention baseline?

Source video
factualVerification needed

Jira Agent in Slack triaged 900 pieces of customer feedback with 90% accuracy.

Evidence

Jira team used Jira Agent in Slack to triage 900 pieces of customer feedback with 90% accuracy

Question

How was triage accuracy defined and against what labelled ground truth?

Source video
factualVerification needed

A PM and designer vibecoded a working alpha from scratch before engineers were added, then shifted from rowing to steering.

Evidence

PM and designer vibecoded an alpha from scratch, proving the value before bringing in engineers

Question

Was the vibecoded alpha promoted to production or rebuilt, and what was the total cost to reach a shippable state?

Source video
factualVerification needed

Fixing design bugs with Figma MCP plus coding agents resolved 14 design bugs in an hour.

Evidence

Using Figma MCP and coding agents to fix 14 design bugs in an hour

Question

Were the 14 fixes production-quality and individually reviewed, or was this a batch of low-complexity changes?

Source video
opinionVerification not requested

Roles are expanding rather than converging, with jobs overlapping in the AI builder era while the PM job's objective stays the same and only the how changes.

Evidence

Instead of roles going away, jobs are overlapping in the AI builder era.

Source video
comparativeVerification needed

Remix with Rovo and Confluence slides shipped in 6 weeks instead of 6 months.

Evidence

Remix with Rovo and Confluence slides launching in 6 weeks instead of 6 months.

Question

Was the 6-month figure a prior estimate or actual historical delivery time for comparable scope?

Source video

Connections

5